kunfang98927/MusicGroundingBench-models

MusicGroundingBench models

LoRA adapters and audio projectors fine-tuned on MusicGroundingBench, a benchmark for note-level temporal grounding and music understanding. These are the fine-tuned models evaluated in the paper Note-Level Temporal Grounding of Musical Concepts in Large Audio-Language Models (arXiv:2608.29480).

• Dataset: https://doi.org/10.5281/zenodo.23377894
• Code and demo: https://github.com/kunfang98927/musicgroundingbench

Built with Meta Llama 3 (the models in Llama-3-8B-MusicGroundingBench/).

Checkpoints

Two backbones, five models each:

• Llama-3-8B-MusicGroundingBench/ — on meta-llama/Meta-Llama-3-8B
• Qwen3.5-9B-MusicGroundingBench/ — on Qwen/Qwen3.5-9B-Base

Folder (in each backbone folder)	Task	Training data	Step (Llama 3)	Step (Qwen3.5)	Paper
2b_grounding/	MusicGroundingBench-2B grounding	grounding training set (33,156 questions)	4000	5185	Table 1, MGBench-2B
2b_understanding_U/	MusicGroundingBench-2B understanding	understanding training set (34,418 questions)	5000	4800	Table 2, "Train on U"
2b_understanding_U05G/	MusicGroundingBench-2B understanding	understanding (34,418) + 50% of grounding (16,578, randomly sampled)	7600	5400	Table 2, "Train on U + 0.5G"
2b_understanding_U10G/	MusicGroundingBench-2B understanding	understanding (34,418) + 100% of grounding (33,156)	8200	5400	Table 2, "Train on U + 1.0G"
3n_grounding/	MusicGroundingBench-3N grounding	grounding questions	2000	2800	Table 1, MGBench-3N

Steps are optimizer updates; the step given is the checkpoint evaluated in the paper.

How the models work

• Audio input. Frame-level features of the frozen MERT-v1-95M model: the audio is mixed down to mono, resampled from 44.1 kHz to 24 kHz, and the 13 hidden states of MERT are averaged, giving 768-dimensional features at 75 frames per second (at most 750 frames, i.e. 10 s).
• Projector. A two-layer MLP (music_encoder.pt, hidden size 2048) maps the features to the hidden size of the language model; the projected frames are inserted into the prompt in place of audio placeholder tokens.
• Language model. The backbone with LoRA (rank 16, alpha 32, dropout 0.05). Grounding answers are time spans written as From X seconds to Y seconds; ...; understanding answers are short free text.

Files

<backbone>/<model>/
  lora_adapter/          PEFT LoRA adapter (adapter_config.json, adapter_model.safetensors)
                         and the tokenizer with the added special tokens
  music_encoder.pt       audio projector weights (PyTorch state dict)
  training_config.yaml   training settings (data, LoRA, batch size, epochs, selected step)
  trainer_state.json     training log (losses, validation losses)

The inference code (prompt construction, feature extraction, generation) will be released in the GitHub repository. Model outputs can be scored with the evaluation code that is already there (evaluation/).

License

See LICENSE.md.

• Llama-3-8B-MusicGroundingBench/: derived from Meta Llama 3 and distributed under the Meta Llama 3 Community License (full text in Llama-3-8B-MusicGroundingBench/LICENSE, required notice in Llama-3-8B-MusicGroundingBench/NOTICE); use is also subject to the Meta Llama 3 Acceptable Use Policy. The base model must be obtained from Meta under the same license.
• Qwen3.5-9B-MusicGroundingBench/: Apache License 2.0, the license of the base model.

Citation

@article{fang2026notelevel,
  title   = {Note-Level Temporal Grounding of Musical Concepts in Large Audio-Language Models},
  author  = {Fang, Kun and Wang, Ziyu and Fujinaga, Ichiro},
  journal = {arXiv preprint arXiv:2608.29480},
  year    = {2026},
  doi     = {10.48550/arXiv.2608.29480}
}